Feature Information Encoding with Transform, Quantization, and Rearrangement

The patent applies domain transformation (e.g., DCT, DST, or learned transforms) and quantization as preprocessing steps before encoding. By performing these transformations in advance, the encoder works with transformed and quantized coefficients that are more amenable to compression, thereby improving encoding efficiency while managing complexity through standardized transform operations.

The patent changes the domain representation of the data by applying linear transforms (such as discrete cosine transform or discrete sine transform) and quantization operations. These parameter changes convert the original image and feature map data into a different domain where the encoding can be performed more efficiently, trading off some processing complexity for improved compression performance.

The patent extracts feature information from intermediate layers of a neural network and performs preprocessing (domain transformation, quantization, rearrangement) before the final encoding stage. By preparing the feature information in advance, the system improves the quality of data available for machine-learning tasks while managing encoding time through efficient transform and quantization operations.

The patent processes different components (image information and feature map information) separately through the transformation and quantization pipeline, then combines them for final encoding. This segmentation allows for optimized processing of each component type, improving overall task performance while managing encoding time through parallel or sequential processing of distinct data streams.